We identified that the solution is to build AIs which do not decide to act on a snapshot of data. Rather they should first determine whether or not they have enough information to act — and what data they need in order to act. This means that AIs in Healthcare, to be successful, must be inextricably linked to their own data sources: and those data sources must be continuous.
Senti uses continuous bioacoustic monitoring to determine the best point to act in the course of a respiratory illness. But were there other choices? Who else is working on this?
Before we can answer these questions, first, let’s return to the basis of breathlessness.
Jess, a 27-year-old woman, goes to drop off a loaf of banana bread with a friend to celebrate lockdown coming round once again. Her friend’s dog, a golden retriever, suddenly jumps up to greet her. The banana bread falls to the ground; rendered inedible. Jess and her friend exchange apologetic pleasantries (this is the UK, after all), but Jess is starting to feel hot and tight-chested. She coughs (into her arm, of course) starting to panic: could this be covid? Moments later, she’s short of breath, unable to talk in full sentences. The golden retriever, realising something is wrong, calls an ambulance*.
*Note: some elements of this story might have been changed to protect the identity of the individuals involved.
What’s happening here?
- The first stage of any condition is the occurrence of some pathological process. For example, bacteria colonise your respiratory tract in pneumonia; inflammatory cells and molecules migrate into the airways in Asthma; weakening of the left ventricle of your heart causes increased fluid pressure in the pulmonary circulation in heart failure.
- The second stage is your body’s subsequent response to this process. Your immune system combating the bacterial infection; your airways constricting in Asthma; fluid moving out of the pulmonary circulation (and into the lung) in heart failure.
- The third stage is your body’s adaptation/compensation to these processes: invariably, in diseases affecting the lungs, this means increasing your breathing rate in an attempt to increase oxygenation.
- The final stage is your body moving into a state of decompensation. Again, invariably, in lung disease, this is a failure of ventilation seen as falling oxygen levels within your body.
Following exposure to these antigens (in this case, the fur from the golden retriever), Jess’ immune system kicks in. Inflammatory cells (particularly eosinophils) migrate into the airways to fight the perceived threat. Cells lining the airway release chemical signals such as histamines. This process can take minutes to hours.
Jess’ attack quickly progresses to stage 2. Jess’s airways are hypersensitive; so these immune signals trigger the airway smooth muscles to constrict, narrowing the small airways. This causes Jess to feel tight-chested. As airway constriction progresses, Jess moves into stage 3. The narrowed airways are now significantly reducing the amount of oxygen delivered into the lungs with each breath. To compensate, she starts to breathe rapidly, maintaining oxygen delivery to the body.
It is vital that any AI monitoring system leverages data which correlates to an earlier part of the process, prior to symptoms developing. Otherwise, it is impossible to create clinically meaningful screening. Patients know that they have developed symptoms: they know they need to seek medical attention.
The futility of AI-Triage systems: a problem of data
AI triage systems currently attempt to utilise symptoms in order to make decisions about whether and where a patient should seek help. However — even in A&E, when I’m 99% confident (from a detailed history: a lot more data than a triage-AI has) that a patient could have waited to see their GP, or even managed their condition at home, I’m still ordering blood tests, X-Rays, and conducting a thorough physical examination. There’s a reason for this: I will not be sending 1 in 100 of my patients (the other 1%) home to die or deteriorate.
Okay — some will argue that we over-investigate, and perhaps many of these investigations could have been avoided. But, regardless, there isn’t a single patient who will not have the benefit of a 20-minute consultation from someone who has been studying medicine for over 5 years. There is no doctor, on the front door of an A&E, turning patients away after talking to them for a couple of minutes: we know this would not be safe.
Healthcare in nuanced. Tom is a tall 32-year-old; he’s usually fit and well — in fact, that morning, he was in the gym as usual. Just as he’s lifting his personal best in weights, he feels some sudden pain in his chest. Tom stops and rests, but the pain fails to subside. It transpires that Tom failed to warmup: he has obviously pulled a muscle. But Tom is worried about the pain; he also wonders — does he feel a little breathless? Tom attends A&E.
The A&E Registrar sighs as he picks up his notes: another twenty-to-thirties young guy with chest pains and some anxiety. Looking at the triage information: chest pain when lifting weights, the diagnosis is immediately apparent. Another pulled muscle. Nothing broken. On examining Tom, there is even some tenderness over the pectoralis muscles, all but confirming the diagnosis. Of course; Tom is sent for a chest X-Ray. There’s a very good reason for this: the X-Ray demonstrates a pneumothorax. That is, a hole in the lungs which causes air to fill the chest cavity. This squashes the lungs and can lead to severe breathlessness. If untreated, the air can continue to acumulate around (rather than inside) the lungs, preventing adequate oxygenation and eventually squashing the heart prevetnting it from beating.
This is, of course, very rare. But it is also very very bad. Tom has an undiagnosed condition called Marfan’s syndrome which can predispose to pneumothroax: but the only clue in the history is how tall Tom is. Clearly, this is a detail which will be missed by most symptom tracking AIs or telephone consultations.
We can just about triage safely in A&E between waiting in the waiting room and waiting in a cubical under observation. People cannot self-triage safely at home: not if we have any kind of ambition for a health service which saves more lives and delivers higher quality care than the (certainly life-saving and high quality) service which we currently have. AI cannot self-triage safely either — not based on history alone. Betwixt decisions leading to disaster or tragedy, symptom-based triage AIs are futile.
I’m often asked to explain what differentiates Senti from Hexoskin, Prevayl, and other autonomous/smart vital sign monitoring solutions. Don’t misunderstand me — these look like great tools for fitness and wellbeing: some might even help improve in-hospital processes (i.e monitoring patients who we already know are unwell, in need of hospital treatment, and whom we want to monitor for signs of recovery or further deterioration). But, when it comes to revolutionising the way in which we seek care and act earlier in disease processes, the differentiation is clear: these devices can only ever tell you what the patient already knows —either they are unwell and feel unwell, or, they feel okay and the machine can tell no different.
The benefits of pre-symptomatic monitoring
Senti, on the other hand, by tracking bioacoustic markers, detects the body’s response to a pathological process. We can detect illness before it manifests in symptoms*. And, through continuous monitoring of these markers, Senti can check its own answers — and, rather than needing to ask forgiveness if it makes a mistake (and a patient unexpectedly deteriorates), Senti instead will ask for an ambulance.
*The proof will be in the outcomes of our clinical trials over the next 12 months; but this hypothesis follows analytically from measuring markers arising from the second stage
First, your body is quite adept at managing these underlying processes to a degree. There are plenty of colds and bacterial chest infections which will simply get better by themselves. If we act on these markers, we risk over-medicalising wellbeing. Treatments come with risks and consequences: the looming crisis of antimicrobial resistance is a manifestation of this pathology.
John, a 74-year-old gentleman with Chronic Obstructive Pulmonary Disease (COPD) suffers from frequent flareups of his condition. In COPD, airway secretions (mucus) can become very thick and sticky. This is difficult to clear and stagnant mucus can become a focus for an infection to take hold. These infections trigger similar constriction of the small airways to asthma (possibly as a mechanism for trying to clear these secretions), leading to breathlessness.
However, the point here is that there is often this colonisation of bacteria: colonisation does not necessitate infection. Many attacks can be cleared simply using steroids which actually dampens the immune response. Sounds counter-intuitive? But the key problem in COPD is the immune response to these bacteria, as this leads to airway constriction: not the presence of these bacteria themselves. If we treat every time we detect bacteria, we will disrupt the normal colonisation of bacteria which should live in your airways and, more importantly, surviving bacteria will become resistant to the antimicrobials which we are using. Our fear is that when a more substantial infection takes hold, we will have no treatment options due to this resistance there will be nothing we can do.
If we do not give AI the choice to gather more data, then we prevent AI from being able to make clinically meaningful or safe decisions. Ultimately, the practice of medicine is about gathering enough data to make the right decisions. If we do not think we have enough data ourselves, then we order investigations or refer for specialist opinions. This is how medicine works. To prevent AI from having the same options is to mistake mathematics for magic.
There is a final piece to consider: direct competitors who are also conducting the same bioacoustic monitoring. Some of these — such as MamaOpe, who are doing some fantastic work solving the problem of access to pneumonia diagnosis in Uganda — only capture a snapshot of data. To be clear, given the use case, this is absolutely appropriate. A snapshot will allow limited diagnosis via AI, but this technology is unlikely to improve substantially against a clinician reading a chest X-Ray. It will help to improve the diagnosis and treatment of pneumonia in countries lacking access to Chest X-Rays, but it will not enable precision and predictive medicine for respiratory illnesses.
Other competitors, such as Strados labs and the Addam from HealthCare originals understand the importance of long term monitoring — but are only monitoring the lungs in one place. This does unfortunately give them a differentiation problem (distinguishing one condition from another). If you assume that the patient’s Asthma is definitely the route cause of any change in that person’s lung sounds, this technology will, in the lab, appear effective. But, again, unfortunately, the world of healthcare is messy, and these assumptions can be catastrophic. If your sensor is placed in the upper left of the chest, what hope do you have of detecting the pneumonia brewing in the right lower lobe?
Whilst Strados might have detected Jess’ rapidly evolving asthma attack, it would fail to determine whether John was suffering from an infective or non-infective attack of COPD and could entirely miss Tom’s pneumothorax.
Senti is currently recruiting pilot sites for our clinical trials. To learn more, reach out to us here: https://senti.care.